Repository navigation
[Klaud Cold] Update GLM-5.2 FP4 GB300 Dynamo-SGLang AgentX agg image to nightly-dev-20261021 / [Klaud Cold] 将 GLM-5.2 FP4 GB300 Dynamo-SGLang AgentX agg 镜像更新至 nightly-dev-20261021 - #4449
Conversation
…261021-c1c287a4 Move glm5.2-fp4-gb300-dynamo-sglang-agentic-agg to the digest-pinned SGLang nightly c1c287a4 and record the bundled ai-dynamo version. 将 glm5.2-fp4-gb300-dynamo-sglang-agentic-agg 更新到按摘要固定的 SGLang nightly c1c287a4 镜像,并记录其内置的 ai-dynamo 版本。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thanks for the contribution!
中文感谢你的贡献!
|
|
Initial attempt · Queued · Run 40006929339 / attempt 1 · 2026-10-23 18:53 UTC
Note: All rows: pending. Next: Wait for the smoke benchmarks and the representative eval of run 40006929339, which measures branch commit 4551c5d. 中文初次尝试 · 排队中 · Run 40006929339 / attempt 1 · 2026-10-23 18:53 UTC |
|
Initial attempt · Queued · Run 40006929339 / attempt 2 · 2026-10-23 20:01 UTC
Note: All rows: pending. Next: Run attempt 2 has waited about 30 minutes in the low-priority gb300-nv queue with all three jobs still queued; keep waiting for its smoke benchmarks and representative eval. 中文初次尝试 · 排队中 · Run 40006929339 / attempt 2 · 2026-10-23 20:01 UTC |
|
Initial attempt · Queued · Run 40006929339 / attempt 3 · 2026-10-24 00:02 UTC
Note: All rows: pending.
Note: All rows: Δ N/A: no matched eval baseline. Next: Run attempt 3 has waited over two hours in the low-priority gb300-nv queue with both benchmark jobs still queued; keep waiting without cancelling. 中文初次尝试 · 排队中 · Run 40006929339 / attempt 3 · 2026-10-24 00:02 UTC |
|
resumed · The session ended before finishing; session 2 continues it in a new run of its planner. 中文resumed · 会话未完成即已结束;第 2 个会话将在其规划运行的新运行中继续。 |
|
Initial attempt · Passed · Run 40006929339 / attempt 4 · 2026-10-24 02:33 UTC
Note: All rows: Δ N/A: no matched eval baseline. Next: Both smoke benchmarks and the representative gsm8k eval passed, so append the perf-changelog entry and dispatch the final full sweep. 中文初次尝试 · 已通过 · Run 40006929339 / attempt 4 · 2026-10-24 02:33 UTC |
…094c2d-225fe14aaa75010c
…mage update 为 GB300 Dynamo+SGLang agg 镜像更新添加 perf-changelog 条目。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Limit the master-config change to the image so the AgentX point identities match the frozen baseline. 将 GB300 agg 的 router 元数据保留为基线版本,使主配置仅更改镜像,AgentX 点位标识与冻结基线一致。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Final full sweep · Passed · Run 40054839636 / attempt 1 · 2026-10-24 04:11 UTC
Note: All rows: Δ N/A: no matched eval baseline. Next: Mark ready for maintainer review. 中文最终完整 sweep · 已通过 · Run 40054839636 / attempt 1 · 2026-10-24 04:11 UTC |
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40054839636 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40054839636 |
|
validated · Repairs: 1 · Runs: 40006929339, 40054708104, 40054816880, 40054839636 中文validated · 修复次数:1 · 运行:40006929339, 40054708104, 40054816880, 40054839636 |
|
/use 40054839636 |
|
@Klaud-Cold staged run 40054839636. This run remains available across future |
There was a problem hiding this comment.
Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.
Beyond the inline finding, I checked two other things and neither is a problem:
- The sibling glm5.2-fp4-gb300-dynamo-sglang-agentic-disagg row in the same recipe dir is still on the old image. It reads disagg-variants.yaml, not agg-variants.yaml, so the framework-version bump here doesn't change it.
- The old tag in the perf-changelog entry (nightly-dev-20261016-a93f16f0) matches the image this PR removed.
Extended reasoning...
This PR bumps the SGLang image (pinned by digest) for one GB300 agentic agg config, updates the framework versions in the matching agg-variants.yaml recipe, and appends one perf-changelog entry. It touches no security-sensitive surface. One finding about the stale router version is posted inline. This note records only the extra checks that were done and ruled out.
Additional findings (outside the current diff — GitHub can't attach inline comments there):
-
🟡
inferencex-e2e/configs/nvidia-master.yaml— Benchmark results for this config will record the wrong Dynamo router version after merge, 1.6.0.dev20261015 instead of the 1.6.0.dev20261019 the new image ships. The image moved to nightly-dev-20261021-c1c287a4 and the recipe now declaresdynamo: 1.6.0.dev20261019. Butrouter.versionat nvidia-master.yaml:7624 was not updated. Fix: setrouter: { name: dynamo-router, version: "1.6.0.dev20261019" }to match the recipe's frameworks.dynamo. The GB200 sibling keys at nvidia-master.yaml:7448 and :7489 already follow this pattern.Why this was flagged
The PR bumps the image for glm5.2-fp4-gb300-dynamo-sglang-agentic-agg (nvidia-master.yaml:7618). It also bumps agentx/agg-variants.yaml:9 to
dynamo: 1.6.0.dev20261019, which is the Dynamo bundled in the new image. The master row'srouter: { name: dynamo-router, version: "1.6.0.dev20261015" }at nvidia-master.yaml:7624 is left unchanged. srt_slurm/generate.py:53 passespoint["router"]into ROUTER_METADATA. The results processors (results/agentic/init.py:203) write that metadata into each benchmark record. So every run from this PR is labelled with Dynamo 1.6.0.dev20261015 while 1.6.0.dev20261019 actually runs, and the published comparison against the baseline shows no router change when there was one. On the base branch the label matched the bundled Dynamo of nightly-dev-20261016 (per the perf-changelog entry at line 10147). The GB200 sibling keys bumped to 1.6.0.dev20261019 updated their router version together (nvidia-master.yaml:7448, 7489).Verification: Trigger: any run of this config after merge. nvidia-master.yaml:7618 moves the image to nightly-dev-20261021-c1c287a4, but :7624 still has
router: { name: dynamo-router, version: "1.6.0.dev20261015" }, while agg-variants.yaml:9 now declaresdynamo: 1.6.0.dev20261019. On the base, both versions matched. Commit 7cc0765 kept the old version on purpose to match the frozen baseline.
Resolve perf-changelog.yaml by taking main's file and re-appending this PR's entry. 中文:合并 origin/main;perf-changelog.yaml 采用 main 版本,并在末尾重新追加本 PR 的条目。
|
[by Claude Code] The SGLang cookbook update of this PR's AgentX config ( The commands mirror the launch measured in run 40054839636, pinned by this PR's |
|
[by Claude Code] sgl-project/sglang#46807 has been merged. This PR's AgentX config ( The commands mirror the launch measured in run 40054839636, pinned by this PR's |
|
[by Claude Code] The SGLang cookbook update of this PR's AgentX config ( The commands mirror the launch measured in run 40054839636, pinned by this PR's |
|
[by Claude Code] sgl-project/sglang#46948 has been merged. This PR's AgentX config ( The commands mirror the launch measured in run 40054839636, pinned by this PR's |
Goal: Update SGLang image from
lmsysorg/sglang:nightly-dev-20261016-a93f16f0@sha256:0b3e39704480c423d8a00ca15ebee63154defc62944df2d4b11c594563bab457tolmsysorg/sglang:nightly-dev-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb.Baseline: 2026-10-20 ·
lmsysorg/sglang:nightly-dev-20261016-a93f16f0@sha256:0b3e39704480c423d8a00ca15ebee63154defc62944df2d4b11c594563bab457Mean latency · Sources: API 1, API 2, API 3, API 4
Eval: N/A
Stop: Add the
klaud-abortlabel at any time to stop this candidate now; Klaud cancels its runs, closes this PR and does not retry it. Remove the label to let the planner pick it up again.中文
**目标:**将 SGLang 镜像从
lmsysorg/sglang:nightly-dev-20261016-a93f16f0@sha256:0b3e39704480c423d8a00ca15ebee63154defc62944df2d4b11c594563bab457更新至lmsysorg/sglang:nightly-dev-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb。**基线:**2026-10-20 ·
lmsysorg/sglang:nightly-dev-20261016-a93f16f0@sha256:0b3e39704480c423d8a00ca15ebee63154defc62944df2d4b11c594563bab457平均延迟 · 来源: API 1, API 2, API 3, API 4;数值及异常说明见上表。
**停止:**随时添加
klaud-abort标签即可立即停止该候选:Klaud 会取消其运行、关闭此 PR 且不再重试。移除该标签后,规划器会在下次扫描时重新考虑它。